talon

Talon 使用指南

Talon 是面向 AI 应用的多模融合数据引擎，单二进制、零外部依赖，提供 9 大引擎：SQL、KV、Vector、TimeSeries、MessageQueue、Full-Text Search、GEO、Graph、AI。

连接

use talon::Talon;
use talon_ai::TalonAiExt;
let db = Talon::open("./data")?;  // 嵌入式模式，数据目录

多语言 SDK 均通过 FFI 绑定 libtalon，接口模式一致。详见 references/sdk.md。

引擎选择

快速上手

SQL

db.run_sql("CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT, emb VECTOR(384))")?;
db.run_sql("INSERT INTO users VALUES (1, 'Alice', '[0.1, ...]')")?;
db.run_sql("SELECT id, vec_cosine(emb, '[0.1, ...]') AS score FROM users ORDER BY score LIMIT 10")?;

KV

db.kv()?.set(b"user:1", b"Alice", None)?;           // 无 TTL
db.kv()?.set(b"session:x", b"tok", Some(3600))?;     // 1h TTL
let val = db.kv_read()?.get(b"user:1")?;

Vector

db.run_sql("CREATE VECTOR INDEX idx ON docs(emb) USING HNSW WITH (metric='cosine')")?;
let ve = db.vector("idx")?;
let hits = ve.search(&query_vec, 10, "cosine")?;      // (id, score)

AI (Session / Memory / RAG)

let ai = db.ai()?;
ai.create_session("chat-1", BTreeMap::new(), None)?;
ai.append_message("chat-1", &ContextMessage { role: "user".into(), content: "Hi".into(), token_count: Some(1) })?;
let history = ai.get_context_window("chat-1", 4096)?; // 自动截断到 token 预算
ai.store_memory("chat-1", "用户偏好 Rust", &embedding, None)?;
let mems = ai.search_memories("chat-1", &query_emb, 5)?;

跨引擎融合查询

// Hybrid Search: BM25 + Vector (RRF)
let hits = hybrid_search(&store, &HybridQuery {
    fts_index: "articles", vec_index: "emb_idx",
    query_text: "AI database", query_vec: &emb, limit: 10, ..Default::default()
})?;

// GraphRAG: 图遍历 + 向量相似
let hits = graph_vector_search(&store, &GraphVectorQuery {
    graph: "knowledge", vec_name: "embeddings",
    start: root_id, max_depth: 3, direction: Direction::Out,
    query_vec: &emb, k: 10,
})?;

详细 API 参考

按需查阅对应 reference 文件：

• SQL 引擎 (DDL/DML/函数/窗口/CTE/事务): references/sql.md
• KV 引擎 (CRUD/TTL/计数器/扫描/快照): references/kv.md
• Vector 引擎 (HNSW/metadata filter/量化/recommend/discover): references/vector.md
• AI 引擎 (Session/Context/Memory/RAG/Agent/Trace/Intent): references/ai.md
• 其他引擎 (TS/MQ/FTS/GEO/Graph/Fusion): references/more-engines.md
• 多语言 SDK (Go/Python/Node.js/Java/.NET): references/sdk.md

AI 应用最佳实践

RAG 管道

用户问题 → Embedding → hybrid_search(FTS + Vector) → Top-K chunks → LLM Context → 回答

用 ai.store_document() 存储文档分块，ai.search_chunks() 向量检索，ai.search_chunks_hybrid() 混合检索。

Agent 工具缓存

ai.cache_tool_result("weather", &args_hash, result_bytes, Some(3600))?;
if let Some(cached) = ai.get_cached_tool_result("weather", &args_hash)? { return cached; }

对话管理

ai.set_system_prompt("chat-1", "你是一个 Rust 专家")?;
let (prompt, msgs) = ai.get_context_window_with_prompt("chat-1", 4096)?;
// prompt + msgs 直接送入 LLM API

Embedding 缓存

let hash = sha256(text);
if let Some(emb) = ai.get_cached_embedding(&hash)? { return emb; }
let emb = call_openai_embedding(text)?;
ai.cache_embedding(&hash, &emb)?;

执行追踪

ai.log_trace(&TraceRecord {
    run_id: "run-1".into(), session_id: Some("chat-1".into()),
    operation: "llm_call".into(), input: json!({"model": "gpt-4"}),
    output: None, latency_ms: 230, token_usage: Some(150),
})?;
let report = ai.trace_performance_report(Some("chat-1"))?;